unsloth
Qwen2.5 VL 3B Instruct
Qwen2.5 VL 3B Instruct is a 3.09B dense qwen2vl model from unsloth. Configured context length is 128,000 tokens. Quant sizes are the GGUF file bytes published on Hugging Face when a file exists.
Parameters3.1B
Context125K
Attentiongqa
Licensesee the repo
36 layers · 2 KV heads · head dim 128 · GGUF repo · config
Quant file sizes
| Quant | File size | Source |
|---|---|---|
| BF16 | 5.75 GB | Hugging Face file bytes |
| UD-Q8_K_XL | 3.66 GB | Hugging Face file bytes |
| Q8_0 | 3.06 GB | Hugging Face file bytes |
| UD-Q6_K_XL | 2.58 GB | Hugging Face file bytes |
| Q6_K | 2.36 GB | Hugging Face file bytes |
| UD-Q5_K_XL | 2.08 GB | Hugging Face file bytes |
| Q5_K_M | 2.07 GB | Hugging Face file bytes |
| Q5_K_S | 2.02 GB | Hugging Face file bytes |
| UD-Q4_K_XL | 1.82 GB | Hugging Face file bytes |
| Q4_K_M | 1.80 GB | Hugging Face file bytes |
| Q4_K_S | 1.71 GB | Hugging Face file bytes |
| IQ4_NL | 1.70 GB | Hugging Face file bytes |
| IQ4_XS | 1.62 GB | Hugging Face file bytes |
| Q4_1 | 1.86 GB | Hugging Face file bytes |
| Q4_0 | 1.70 GB | Hugging Face file bytes |
| UD-Q3_K_XL | 1.52 GB | Hugging Face file bytes |
| Q3_K_M | 1.48 GB | Hugging Face file bytes |
| Q3_K_S | 1.35 GB | Hugging Face file bytes |
| UD-IQ3_XXS | 1.21 GB | Hugging Face file bytes |
| UD-Q2_K_XL | 1.22 GB | Hugging Face file bytes |
| Q2_K | 2.37 GB | Hugging Face file bytes |
| UD-IQ2_XXS | 0.90 GB | Hugging Face file bytes |
| UD-IQ1_M | 0.82 GB | Hugging Face file bytes |
| UD-IQ1_S | 0.77 GB | Hugging Face file bytes |
On common machines
| Device | Verdict | Quant | Speed |
|---|---|---|---|
| GeForce RTX 4060 Ti 8GB | Tight fit | BF16 | 24–33 |
| GeForce RTX 3060 12GB | Runs great | BF16 | 31–42 |
| GeForce RTX 4060 Ti 16GB | Runs great | BF16 | 24–33 |
| GeForce RTX 4090 | Runs great | BF16 | 86–117 |
| GeForce RTX 5090 | Runs great | BF16 | 152–207 |
| Apple M4 Pro | Runs great | BF16 | 23–32 |
| Apple M5 | Runs great | BF16 | 13–18 |
| CPU only · DDR5-5600 | Runs on CPU | BF16 | 7.6–10 |
Smallest devices that fit it fully
- GeForce RTX 3050 6GB — Runs, UD-Q8_K_XL, 22–30 tok/s
- GeForce RTX 2060 6GB — Runs, UD-Q8_K_XL, 44–59 tok/s
- GeForce GTX 1660 Super — Runs, UD-Q8_K_XL, 44–59 tok/s
- GeForce GTX 1660 Ti — Runs, UD-Q8_K_XL, 37–51 tok/s
- Intel Arc A380 — Runs, UD-Q8_K_XL, 24–33 tok/s
- GeForce RTX 5060 Ti 8GB — Tight fit, BF16, 38–52 tok/s
Run it
Commands for the GeForce RTX 4060 Ti 16GB estimate (BF16).
llama.cpp
llama-server -hf unsloth/Qwen2.5-VL-3B-Instruct-GGUF:BF16 -c 8192 -ngl 36
Ollama
ollama run qwen2.5
LM Studio
lms get unsloth/Qwen2.5-VL-3B-Instruct-GGUF # Then load quant BF16